Computing with adversarial noise Aram Harrow (UW -> MIT) Matt Hastings (Duke/MSR) Anup Rao (UW)

The origins of determinism Theorem [von Neumann]: There exists a constant p>0 such that for any circuit C there exists a circuit C such that size(C) size(C) * poly log(size(C)) If C is implemented with noise p at every gate, then it will implement C correctly with probability Theorem [Shor 95]: Same is true for quantum computers

Assumptions of FTC 1.parallel operations, i.e. Ω(n) operations per time step. (Assuming that NOP is noisy.) 2.irreversibility / ability to cool bits i.e. Ω(n) fresh 0 bits per time step 3.i.i.d. noise model decaying correlations: e.g. Pr(bits flipped in set S) exp( -c |S|) Q: How far can we relax this assumption?

Adversarial noise Computation model is deterministic circuits, with constant fan-in and fan-out. Number of bits (n) is either static or dynamic. In every time step, the adversary chooses np bits to flip. Input and outputs are assumed to be encoded in a linear-distance ECC. Goals:memorycomputation

Existing constructions fail von Neumann FT (with any code) cannot store ω(1/p) bits. Bits arranged in k-D fails for any constant k. More generally, if gates are constrained to be local w.r.t. any easy-to-partition graph, then memory is impossible

Results GoalNoise rateSize overhead memoryconstant computationconstantexp(o(s)) computation(1/log(s)) O(1/ δ ) s 1+ δ quantum memoryconstantconjectured impossible

The 1->n repetition code or Repeatedly: Partition into random blocks of size 3 and majority vote Claim: Starting with an e 0.1 fraction of errors, we replace this with e 0.9 e + O(p)

Computation with the repetition code If p » 1/k, then k bits can be encoded. Gates can be performed transversally. Example: (a n, b n ) (a n, (a© b) n )

The Hadamard code Definition: k -> n=2 k bits Example: x1x2x3x1x2x3x1x2x3x1x2x3 0x1x2x1©x2x3x1©x3x2©x3x1©x2©x30x1x2x1©x2x3x1©x3x2©x3x1©x2©x30x1x2x1©x2x3x1©x3x2©x3x1©x2©x30x1x2x1©x2x3x1©x3x2©x3x1©x2©x3 Parallel extraction: Can obtain O(n) copies of any bit. x1x1x1x1 © x1x1x1x1 © x1x1x1x1 © x1x1x1x1 © To replace x 1 with f(x 2,x 3 ): 1.Extract x 1, x 2, x 3 into repetition codes 2.Compute x 1 ©f(x 2, x 3 ) transversally 3.XOR this with the appropriate locations in the code

Locally correctable codes Definition: Given a codeword corrupted in a δ fraction of positions, there is a randomized method to recover any coordinate of the original codeword, using q queries and giving a wrong answer with probability ρ. Theorem: Any systematic LCC can be used to make a circuit FT against adversarial noise. Conversely, in any scheme capable of protecting arbitrary circuits against adversarial noise, the input encoding is a LDC. Parameters: q,δ,ρ constant: k bits into exp(o(k)) bits q=log c+1 (k), δ,ρ constant: k bits into k 1+1/c+o(1) bits

FT memory x1x1 x2x2 x3x3 x4x4 x5x5 x6x6 x7x7 x 1 © x 5 © x 6 = 0 x 2 © x 4 © x 7 = 0 x 1 © x 3 © x 7 = 0 LDPC codes Iterative decoding Flip variables who have a majority of unsatisfied neighbors. Key properties 1. Can perform one round of decoding in constant depth. 2. Maps error rate e to e 0.9 e + O(p). 3. Good code (i.e. k/n = Ω(1))

Quantum computing? Immediate difficulties 1. No repetition code: |Ãi |Ãi ­ |Ãi ­ |Ãi is unphysical no analogue of majority-voting correction cannot correct small errors 2. Parallel extraction impossible / no LDCs Quantum states n qubits described by a unit vector in

Stabilizer (linear) QECCs n-qubit Pauli matrices P 1 ­ … ­ P n, for P 1,…, P n 2 {I, X, iY, Z} stabilizer code Given commuting Pauli matrices {s 1, …, s n-k } define the code space V = {|Ãi : s i |Ãi = |Ãi for i=1,…,n-k} rate = (log dim V)/n = k/n distance Let S = and N(S) = {P : sP=Ps for all s2S} Distance = min weight of an element of N(S)\S

Related quantum open question Do there exist stabilizer codes of constant weight and linear distance? Do there exist stabilizer codes of constant weight and linear distance? The only known linear-distance codes have linear-weight generators; the only known constant-weight codes have distance O(n 1/2 log(n)). The only known linear-distance codes have linear-weight generators; the only known constant-weight codes have distance O(n 1/2 log(n)). Homology codes appear unpromising. Homology codes appear unpromising.

Open questions QC with adversarial noise? QC with adversarial noise? Can classical FTC be made more efficient? Can classical FTC be made more efficient? Even against i.i.d. noise, is linear overhead possible? Even against i.i.d. noise, is linear overhead possible? Is code deformation of classical codes possible? What if we had a quantum computer? Is code deformation of classical codes possible? What if we had a quantum computer?